A Qualitative Study of Co-Creation, Communication, Flow, Trust in Vibe Coding (arxiv.org)

This paper presents the first systematic qualitative study of "vibe coding"—the emergent natural-language programming paradigm popularized by Andrej Karpathy—based on over 190,000 words from semi-structured interviews, Reddit threads, and LinkedIn posts. The authors build a grounded theory describing vibe coding as iterative, conversational co-creation with an AI assistant that privileges flow, experimentation, and rapid feedback over heavy upfront specification. Key findings show that developers experience increased flow and enjoyment when the AI supports a conversational co-authoring role; trust in the model governs movement along a continuum from full delegation to tight co-creation. The study surfaces concrete technical pain points and implications for AI/ML tooling: difficulties in precise specification, intermittent reliability and debugging challenges, latency that breaks flow, increased code-review burden, and collaboration friction. The paper also catalogs emerging best practices developers use to mitigate these issues. For the AI/ML community, the work highlights priorities for future research and product design—improving model reliability and explainability, interactive debugging and specification tools, latency reduction, and workflows that preserve auditability and team collaboration—to make vibe coding safe, scalable, and productive.